An Empirical Study on Numerical Bugs in Deep Learning ProgramsVirtual
The task of a deep learning (DL) program is to train a model with high precision and apply it to different scenarios. A DL program often involves massive numerical calculations. Therefore, the robustness and stability of the numerical calculations are dominant to the quality of DL programs. Indeed, numerical bugs are common in DL programs, producing NaN (Not-a-Number) and INF (Infinite). A numerical bug may render the DL models inaccurate, causing the DL applications unusable. In this work, we conduct the first empirical study on numerical bugs in DL programs by analyzing the programs implemented on the top of two popular DL libraries (i.e., TensorFlow and Pytorch). Specifically, We collect a dataset of 400 numerical bugs in DL programs. Then, we classify these numerical bugs into 9 categories based on their root causes and summarize two findings. Finally, we provide the implications of our study on detecting numerical bugs in DL programs.
Thu 13 OctDisplayed time zone: Eastern Time (US & Canada) change
10:00 - 12:00 | Technical Session 21 - SE for AI IIResearch Papers / Late Breaking Results / NIER Track / Journal-first Papers at Banquet B Chair(s): Andrea Stocco Università della Svizzera italiana (USI) | ||
10:00 20mResearch paper | DeepPerform: An Efficient Approach for Performance Testing of Resource-Constrained Neural Networks Research Papers Simin Chen University of Texas at Dallas, USA, Mirazul Haque UT Dallas, Cong Liu University of Texas at Dallas, USA, Wei Yang University of Texas at Dallas | ||
10:20 10mPaper | Prototyping Deep Learning Applications with Non-Experts: An Assistant Proposition Late Breaking Results Gustavo Rodrigues dos Reis Rodrigues dos Reis, Adrian Mos NAVER LABS Europe, Cyril Labbé LIG - UGA, Mario Cortes Cornax LIG - UGA | ||
10:30 20mResearch paper | Boosting the Revealing of Detected Violations in Deep Learning Testing: A Diversity-Guided MethodVirtualACM SIGSOFT Distinguished Paper Award Research Papers Xiaoyuan Xie School of Computer Science, Wuhan University, China, Pengbo Yin School of Computer Science, Wuhan University, Songqiang Chen School of Computer Science, Wuhan University | ||
10:50 20mPaper | Faults in Deep Reinforcement Learning Programs: A Taxonomy and A Detection ApproachVirtual Journal-first Papers Amin Nikanjam École Polytechnique de Montréal, Mohammad Mehdi Morovati École Polytechnique de Montréal, Foutse Khomh Polytechnique Montréal, Houssem Ben Braiek École Polytechnique de Montréal Link to publication DOI Authorizer link | ||
11:10 20mResearch paper | Towards Understanding the Faults of JavaScript-Based Deep Learning SystemsVirtual Research Papers Lili Quan Tianjin University, Qianyu Guo College of Intelligence and Computing, Tianjin University, Xiaofei Xie Singapore Management University, Singapore, Sen Chen Tianjin University, Xiaohong Li TianJin University, Yang Liu Nanyang Technological University | ||
11:30 10mVision and Emerging Results | An Empirical Study on Numerical Bugs in Deep Learning ProgramsVirtual NIER Track Gan Wang , Zan Wang Tianjin University, China, Junjie Chen Tianjin University, Xiang Chen Nantong University, Ming Yan College of Intelligence and Computing, Tianjin University | ||
11:40 20mResearch paper | Toward Improving the Robustness of Deep Learning Models via Model TransformationVirtual Research Papers Yingyi Zhang College of Intelligence and Computing, Tianjin University, Zan Wang Tianjin University, China, Jiajun Jiang Tianjin University, Hanmo You College of Intelligence and Computing, Tianjin University, Junjie Chen Tianjin University | ||